computer-use-agents

Fail

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTION
Full Analysis
  • COMMAND_EXECUTION (HIGH): The skill implements a bash tool (BetaToolBash20241022) and uses the pyautogui library to provide direct, programmatic control over the operating system GUI and shell environment.
  • DATA_EXFILTRATION (MEDIUM): The 'capture_screenshot' method captures the entire desktop state as base64-encoded image data, which is then processed by an external LLM, leading to the exposure of any sensitive information visible on the screen.
  • REMOTE_CODE_EXECUTION (HIGH): The inclusion of 'bash' and 'str_replace_editor' tools enables an AI model to write, modify, and execute arbitrary scripts on the underlying host or container.
  • INDIRECT_PROMPT_INJECTION (LOW): As a vision-language agent, the skill is vulnerable to instructions embedded in screenshots (e.g., malicious text on a website) that could hijack the agent's logic. (Evidence: Ingestion point in capture_screenshot; Capabilities: bash/pyautogui; No input sanitization or boundary markers identified).
  • EXTERNAL_DOWNLOADS (SAFE): The provided Dockerfile uses standard, trusted repositories for Ubuntu and Python package installations.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 17, 2026, 06:26 PM